The invention relates to active acoustic attenuation systems, and provides a system for cancelling undesirable output sound. The system adaptively models and compensates for feedback sound, and also provides adaptive on-line modeling and compensation of the effects of the error path and cancelling speaker.
Prior feedback cancellation systems use a filter to compensate for feedback sound from the speaker to the input microphone. It is desirable that this filter be adaptive in order to match the changing characteristics of the feedback path. Prior systems will successfully adapt only for broad band noise input signals because the system input is uncorrelated with the output of the feedback cancellation filter. Uncorrelated signals average to zero over time. However, if the input noise contains narrow band noise such as a tone having a regular periodic or recurring component, as at a given frequency, the filter output will be correlated with the system input and will not converge. The filter may thus be used adaptively only in systems having exclusively broad band input noise.
Most practical systems, however, do experience narrow band noise such as tones in the input noise. The noted filter cannot be adaptively used in such systems. To overcome this problem, and as is known in the prior art, the filter has been pre-trained off-line with broad band noise only. This pre-adapted filter is then fixed and inserted into the system as a fixed element which does not change or adapt thereafter.
A significant drawback of the noted fixed filter is that it cannot change to meet changing feedback path characteristics, such as temperature or flow changes in the feedback path, which in turn change the speed of sound. During the pre-training process, the filter models a pre-determined set of given parameters associated with the feedback path, such as length, etc. Once the parameters are chosen, and the filter is pre-adapted, the filter is then inserted in the system and does not change thereafter during operation. This type of fixed filter would be acceptable in those systems where feedback path characteristics do not change over time. However, in practical systems the feedback path does change over time, including temperature, flow, etc.
It is not practical to always be shutting down the system and re-training the filter every time the feedback path conditions change, nor may it even be feasible where such changes occur rapidly, i.e., by the time the system is shut down and the filter re-trained off-line, the changed feedback path characteristic such as temperature may have changed again. For this reason, the above-noted fixed filter is not acceptable in most practical systems.
There is thus a need for adaptive feedback cancellation in a practical active acoustic attenuation system, where the characteristics of the feedback path may change with time. A system is needed wherein the feedback is adaptively cancelled on-line for both broad band and narrow band noise without dedicated off-line pre-training, and wherein the cancellation further adapts on-line for changing feedback path characteristics such as temperature and so on.
Co-pending Ser. No. 777,928, filed Sept. 19, 1985, and assigned to the same assignee, discloses a system wherein the feedback is adaptively cancelled on-line for both broad band and narrow band noise without dedicated off-line pre-training, and wherein the cancellation further adapts on-line for changing feedback path characteristics such as temperature.
Co-pending application Ser. No. 777,825, filed Sept. 19, 1985 and assigned to the same assignee, discloses an improved system additionally providing adaptive on-line compensation of the error path between the cancelling speaker and the output. The characteristics of the cancelling speaker are assumed to be relatively constant or to change only slowly relative to the overall system and relative to the feedback path from the cancelling speaker to the input and relative to the error path from the cancelling speaker to the output. While the sound velocity in the feedback path and in the error path may change according to temperature, etc., the characteristics of the cancelling speaker change only very slowly relative thereto. The speaker is thus modeled off-line and calibrated, and assumed to be fixed or at least change only very slowly relative to the other system parameters, especially temperature and flow rate.
The present invention provides a further improved system affording better performance, including adaptive on-line modeling of both the error path and the cancelling speaker, without dedicated off-line pre-training.
The noted co-pending applications provide a technique for active attenuation that effectively solves the problem of acoustic feedback from the secondary sound source cancelling speaker to the input microphone. This technique utilizes a recursive least mean squares RLMS algorithm to provide a complete polezero model of the acoustical plant. An error signal is used to adapt the coefficients of the RLMS algorithm model in such a manner as to minimize the residual noise.
If the speaker transfer function is not to be assumed fixed, or if a lower grade or quality speaker is desired for cost reduction, then both the error path transfer function and speaker transfer function must be compensated for in the algorithm model. Widrow, Adaptive Filters, "Aspects of Network and System Theory", R. E. Kalman and N. Declaris, EDS., New York, Holt, Rinehart and Winston, 1971, has shown that the LMS algorithm can be used with a delayed error signal if the input to the error correlators is also delayed. Similarly, Morgan, "Analysis of Multiple Correlation Cancellation Loop With a Filter in the Auxiliary Path", IEEE Transactions Acoustics, Speech, Signal Processing, Vol. ASSP-28 (4), pp. 454-467, 1980, has noted that the LMS algorithm can be used with a transfer function, such as that due to the speaker, in the auxiliary path if either this transfer function is also inserted in the input to the error correlators or if an inverse transfer function is added in series with the original. Burgess, "Active Adaptive Sound Control in a Duct: A Computer Simulation", Journal of Acoustic Society of America, 70 (3), pp. 715-726, 1981, has discussed similar results when both auxiliary path and error path transfer functions are present.
In an active sound attenuation system using the RLMS algorithm, if both the speaker transfer function S and the error path transfer function E are known, their effect on the convergence of the algorithm may be corrected through either the addition of S and E in the input lines to the error correlators or the addition of the inverse transfer functions S.sup.-1 and E.sup.-1 in series in the error path. Thus, it is necessary to obtain either direct or inverse models of S and E.
Poole et al, "The Implementation of Digital Filters Using a Modified Widrow-Hoff Algorithm for the Adaptive Cancellation of Acoustic Noise", Proceedings ICASSP 84, pp. 21.7.1-21.7.4, 1984, and Warnaka et al U.S. Pat. No. 4,473,906, have described a system using the LMS algorithm in which the delayed adaptive inverse modeling procedure of Widrow et al, "Adaptive Control by Inverse Modeling", Proceedings of 12th Asilomar Conference on Circuits, Systems and Computers, Pacific Grove, Calif. Nov. 6-8, 1978, pp. 90-94, is used to obtain an off-line model of the delayed inverse transfer function models .DELTA.S.sup.-1 E.sup.-1. As noted above, this approach then requires the addition of delay .DELTA. to the input to the error correlators of the LMS algorithm. The above noted co-pending application Ser. No. 777,825, filed Sept. 19, 1985, describes a three microphone system using the RLMS algorithm in which the error plant is modeled on-line using either a direct or inverse model while the speaker is modeled off-line.
In the present invention, the speaker and the error path are modeled on-line. The system functions adaptively in the presence of acoustic feedback, and non-ideal speaker and error path transfer functions. The system responds automatically to changes in the input signal, acoustic plant, error plant and speaker characteristics.
There are two basic techniques available for use in system modeling. The direct model approach places the adaptive model in parallel with the speaker. The impulse response of the model is the same as that of the speaker. The inverse model approach places the adaptive model in series with the speaker. The impulse response of the model represents the delayed inverse response of the speaker. Either approach can be used off-line to determine SE or .DELTA.S.sup.-1 E.sup.-1 for use in the RLMS algorithm as noted above. However, on-line measurements are complicated by the fact that in addition to the model output exciting the speaker S, the plant output is also present at the input to the error path E. The speaker transfer function cannot be determined in this case unless the plant noise, which is correlated with the model output, is removed. The model output or a training signal can be used to determine SE off-line.
The present invention provides a new technique and system for on-line modeling of S and E. An uncorrelated auxiliary random noise source is used to excite the speaker and the error path. The noise level emanating from the speaker will ultimately become the residual noise of the system. A direct adaptive model is used to obtain coefficients describing S and E that can be used in the input lines to the error correlators for the primary RLMS algorithm in the preferred embodiment. The amplitude of the auxiliary uncorrelated noise source is kept very low so that the final effect on the residual noise is small. The plant output noise and the model output are not present at the input to the adaptive SE model and so will not affect the final values of the model weights. The auxiliary noise source is placed following the summing junction of the RLMS algorithm and ensures that the added noise passes through both the electro-acoustic feedback path as well as the recursive loop in the RLMS algorithm and the feedback noise is cancelled as the algorithm converges.
The uncorrelated random auxiliary noise source is independent of the input signal and ensures that the speaker and error path will be correctly modeled. The signals from the plant output and the model represent noise on the plant side of the speaker/error path modeling system and will not affect the weights of the direct LMS model used to determine SE. Copies of this model are provided in the input lines of the error correlators.
It has been found that the use of a delayed adaptive inverse model .DELTA.S.sup.-1 E.sup.-1 will result in decreased performance since the plant noise due to the plant output and model output also appears at the input to the adaptive filter. Thus, the auto-correlation function of the filter input is adversely affected, and the filter weights are modified, Widrow and Stearns, Adaptive Signal Processing, Englewood Cliffs N.J. Prentice-Hall, Inc., 1985, pp. 196, 197, 222, 223. If the plant noise is large enough, the adaptive model may fail to converge. Thus, the delayed adaptive inverse approach requires a much larger amplitude noise source, which increases the residual noise and decreases overall system quieting.
In a direct model system, SE, the plant noise does not affect the final weights in the adaptive model. In addition, the convergence of the SE model is assured as long as the initial amplitudes are within the dynamic range of the system. Thus, with SE acccurately determined, the overall system model will converge, resulting in minimum residual noise. The algorithm properly converges for either narrow band or broad band input signals. The coefficients of the SE model properly describe the SE path, and the coefficients of the overall system model properly describe the plant P, the feedback path F, the error path E, and the speaker S. The invention provides a complete active attenuation system in which acoustic feedback is modeled as part of the adaptive filter, and in which the effects of the sound source and the error path transfer functions are adaptively modeled on-line through the use of a second algorithm that uses a separate low level random auxiliary noise source to model the sound source and error path which the system is operating.